770

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 20, NO. 3, MAY 2012

Comparing Fuzzy and Intelligent PI Controllers in Stop-and-Go Manoeuvres

Vicente Milans, Jorge Villagr, Jorge Godoy, and Carlos Gonzlez, Member, IEEE
is one of the most challenging topics of large city trafc management. These kinds of system are known in the literature as stop-and-go systems [3]. They deal with the vehicle in urban scenarios with frequent and sometimes hard braking and acceleration. The main idea of these control systems is to regulate the vehicle around the well-known 2-s headway rule, which attempts to maintain a distance proportional to the human reaction time (approximately 2 s) [4]. Some approaches have tried to reproduce human behavior with deterministic models in order to achieve smooth control actions. Unfortunately, this kind of strategy may not necessarily lead to safe operation. In [5], Martinez and de Wit proposed a nonlinear reference model taking into account safety and comfort specications in an intuitive way. However, their approach assumes that the reference acceleration generated by the dynamic inter-distance (or gap-distance) model is applied instantaneously to the following vehicle. Since this assumption is hardly ever satised in real urban situations, an advanced feedback controller should be introduced to cope with vehicle nonlinearitiesespecially in brake and engine dynamics at low speedand environmental disturbances. Different approaches have been proposed to tackle the actuators nonlinear dynamics. Input/output linearization [6], fuzzy logic [7], [8], and sliding mode control [9], [10] have been used to deal with engine control. Feedback linearization [11] and sliding modes [12] have also been implemented to control a nonlinear brake model. However, most of these approaches require precise models, so that any parameter variation during the vehicles lifetime may lead to deteriorating performance, or even to unstable behavior. The present work is an attempt to nd an engine/brake control algorithm that yields the expected reference speed and acceleration of the trailing vehicle, while keeping a reference distance from the leading vehicle. Moreover, the control law will have to be robust to measurement noise, unmodeled (brake and engine) dynamics and disturbances (road inclination, aerodynamic forces, and rolling resistance). To that end, two control techniques will be implemented and compared both in simulation and in a commercial vehicle: fuzzy logic and an intelligent proportional integral (PI) controller. The rst of these is one of the class of soft computing techniques [13]. These techniques are recognized as having a strong learning and cognition capability as well as good tolerance to uncertainty and imprecision. Among them, fuzzy logicdeveloped by Prof. L. A. Zadeh in 1965 [14]gives a good approximation to human reasoning, and hence provides an intuitive approach to autonomous control of the nonlinear behavior of commercial vehicles [15]. The second control algorithm uses a novel technique that, based on the classical PI controller structure, is capable of dealing with parameter uncertainties and unmodeled nonlinear dynamics. This control paradigm [16], [17] replaces physical

AbstractThe aim of this work was twofold: on the one hand, to describe a comparative study of two intelligent control techniquesfuzzy and intelligent proportional-integral (PI) control, and on the other, to try to provide an answer to an as yet unsolved topic in the automotive sectorstop-and-go control in urban environments at very low speeds. Commercial vehicles exhibit nonlinear behavior and therefore constitute an excellent platform on which to check the controllers. This paper describes the design, tuning, and evaluation of the controllers performing actions on the longitudinal control of a carthe throttle and brake pedalsto accomplish stop-and-go manoeuvres. They are tested in two steps. First, a simulation model is used to design and tune the controllers, and second, these controllers are implemented in the commercial vehiclewhich has automatic driving capabilitiesto check their behavior. A stop-and-go manoeuvre is implemented with the two control techniques using two cooperating vehicles. Index TermsFuzzy control, intelligent control, nonlinear control, proportional-integral-derivative (PID), road vehicles.

I. INTRODUCTION

AFETY is crucial in the development of autonomous systems in the transportation research eld, with both manufacturers and research groups focusing efforts in this direction. The aim is to develop advanced driver assistance systems (ADAS) that will increase safety in carrying out driving-related tasks. The rst system introduced in commercial vehicles with the potential to inuence trafc safety and trafc ow characteristics was adaptive cruise control (ACC) [1], [2]. ACC is an extension of cruise control (CC)CC allows the driver to set a driving speedin which the vehicle is capable of following a leading car on highways by actions on the throttle pedal. These commercial systems work at speeds greater than 30 km/h. Their main drawback is that they are useless in urban environments. Recently, commercial systems capable of stopping the vehicle when a collision is imminent at speeds below 15 km/h have been developed by car manufacturers, but their dependence on the human driver to restart the vehicle might cause trafc jams. Thus, autonomous intelligent driving in trafc jam conditions
Manuscript received September 15, 2010; revised January 10, 2011; accepted March 21, 2011. Manuscript received in nal form March 24, 2011. Date of publication April 29, 2011; date of current version April 11, 2012. Recommended by Associate Editor A. Alessandri. This work was supported by the Spanish Ministry of Science and Innovation through Research Grant TRANSITO TRA2008-06602-C03 and by the Spanish Ministry of Development through Research Grant GUIADE P9/08. The authors are with the AUTOPIA Program of the Center for Automation and Robotics, Universidad Politcnica de Madrid-Consejo Superior de Investigaciones Cientcas, 28500 Madrid, Spain (e-mail: vicente.milanes@csic.es; jorge.villagra@csic.es; jorge.godoy@csic.es; carlos.gonzalez@csic.es). Color versions of one or more of the gures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identier 10.1109/TCST.2011.2135859

1063-6536/$26.00 2011 IEEE

MILANS et al.: COMPARING FUZZY AND INTELLIGENT PI CONTROLLERS IN STOP-AND-GO MANOEUVRES

771

Fig. 1. Stop-and-Go scheme.

models by local input/output differential equations, valid over short lapses of time. The main advantage of this new approach is that these phenomenological models are merged into a PI transparently, so that an intelligent (hence the name i-PI) term compensates the effects of poorly-known dynamics. In brief, the following issues will be tackled in the present communication. Design and development of two valid solutions for an as yet unresolved issue in the automotive sector: ACC in urban environments at very low speeds. A comparative study of these intelligent control techniques, examining their robustness via a Monte Carlo analysis. Comparison with previously presented solutions [5] to this problem to illustrate the improvements contributed by the present work. Implementation in a commercial cara convertible Citron C3 Pluriel with automated brake and throttleto validate the controllers in a real environment. The rest of this brief is organized as follows. The second section will be devoted to briey presenting the dynamic inter-distance and relative velocity model. In Section III, the design and tuning of the controllers will be presented using a vehicle model. Then the fuzzy and the i-PI controllers will be detailed. Finally, a test of the controllers in a simulation environment will be described, using a Monte Carlo analysis to assess the systems robustness. In Section IV, the two control techniques will be evaluated and compared to a classical PI controller on a real experimental platform, with a focus on comfort and safety aspects. Finally, Section V will present some concluding remarks and a description of future work in this line. II. GENERATION OF THE REFERENCE INTER-DISTANCE AND RELATIVE VELOCITY As mentioned above, the goal of the control strategies will be and , reto use the throttle and brake (control variables spectively) to track as precisely as possible a reference distance between vehicles and a target relative velocity . A reference model proposed by [5] will provide these two variables, and the the trailing car should have to follow the ideal acceleration trajectories of those two reference variables. Note in Fig. 1 that is related to the safe nominal inter-distance the maximum distance at which the control algorithm will be activatedand the critical distance the minimum distance between cars which is only attained when they are stopped. Note also that the dynamic reference model used in the present work will provide a reference inter-distance less than the 2-s headway rule if the allowed maximum acceleration is high enough (for more details, see Fig. 2).

Fig. 2. Comparison of different distance policies: constant headway rule (2 s) and the inter-distance model [5] with different maximum accelerations.

The inter-distance reference model describes the virtual dy(the namics of a vehicle which is positioned at a distance reference distance) from the leading vehicle (1) where is the leading vehicles acceleration and (2) is the trailing acceleration, which is a nonlinear function of the inter-distance and its temporal derivative. in (2), where is the safe nomConsidering inal inter-distance, the control problem is then to nd a suitable , such that all the solutrailing car acceleration , when tions of (1) satisfy the following comfort and safety constraints: , with the minimal inter-distance; , where is the maximum attainable lon gitudinal acceleration; , with a driver desired bound on the jerk. Martinez and de Wit [5] propose the use of a nonlinear where plays the role of a damping model damping constantwhich can be introduced into (1) to give

This equation can be integrated analytically and expressed in as follows: terms of

(3) Note that this reference speed depends upon the leading vehicles speed, the distance , and the parameter , which is in turn an algebraic function of the safety and comfort parameters , and [5]. Fig. 2 shows how inuences the reference inter-vehicle distance. Finally, from (2) the trailing acceleration is (4)

772

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 20, NO. 3, MAY 2012

where the inter-distance evolution is given by the numerical integration of (3). Once the reference inter-distance and its time derivative ( and , respectively) have been generated, the fuzzy and i-PI controllers will seek to ensure that these two variables are tracked as closely as possible. III. DESIGN AND TUNING OF THE CONTROLLERS To design the controllers, a vehicle model was developed which includes the vehicles dynamics at very low speeds so as to constitute a good starting point in tuning the controllers. This section briey describes the model and the design of each of the controllers, a simulation comparing the controllers before their implementation in a real car, and nally Monte Carlo simulations to analyze the robustness to disturbances or parameter uncertainties. A. Vehicle Model The balance of forces along the vehicles longitudinal axis [18] is (5) where is the mass of the vehicle, the longitudinal velocity, and the front and rear longitudinal tyre forces, respecand the front and rear tyre forces due to rolling tively, resistance, the angle of inclination of the road, and the longitudinal aerodynamic drag. The rolling resistance forces are often modeled as linear func, with tions of the normal forces on each tyre, i.e., the rolling resistance coefcient and the vertical load of the vehicle. The aerodynamic forces can be expressed as

TABLE I MODEL PARAMETERS

with , and being the static gain, damping factor, and natural frequency, respectively. Since the braking dynamics is much faster than that of the vehicle, it can be replaced in the vehicle model by an algebraic expression, without loss of realism [21]. The numerical values of the model parameters are given in Table I. B. Fuzzy Controller Fuzzy logic was selected as one of the control techniques to compare because it is a well-tested method for dealing with this kind of system, yields good results, and can incorporate human procedural knowledge into the control algorithms [22]. Also, it allows the designer to partially mimic human driving behavior. The rationale behind the design of the fuzzy controller is to select two errorsdistance and speedas inputs so that the controller can emulate the behavior of a human driver who, in this situation, would control these two parameters. To this end, for each fuzzy variable a central membership function is introduced to dene an area of good performance. Two further membership functions are dened to represent the higher and lower error regions with respect to the target reference. In this sense, to keep a reference distance, a driver does the following: presses the throttle if the distance is greater than the reference distance; presses the brake if the distance is less than the reference distance; lifts the pedal foot off the pedals if the distance is correct. These will also be the actions of our controller since it will try to mimic human behavior. The membership function denition for each of the input variables are shown Fig. 3. The two variables used to perform the control are the SpeedError, dened as the difference between the desired relative velocity between the leading and trailing car , and the DistanceError, deand its actual value ned as the difference between the optimal distance determined using the dynamic inter-distance model generator and the real . Obviously the goal of distance between cars our controller is to perform actions maintaining the values of these two variables close to zero. The choice of the membership functions of each input variable is based on expert drivers experience. In this sense, the goal of our approach to stop-and-go scenarios is based on following both a reference distance and a reference relative velocity. One can easily adjust the fuzzy controller for this purpose

with being the mass density of air, the aerodynamic drag coefcient, the frontal area of the vehicle (the projected area the wind of the vehicle in the direction of travel), and speed. The Pacejka model [19] is used for longitudinal tyre/road interaction forces . The rotation dynamics of each wheel can be expressed as (6) where is the wheels moment of inertia, the angular velocity the applied engine torque, of each wheel, the tyre radius, the brake torque, both applied to each wheels center. and can be expressed in terms of the The total engine torque throttle opening by the expression [20]

where is the gear ratio, is the average wheel speed, is an engine torque parameter, and the maximum torque is . obtained at engine speed Finally, the dynamics between the braking control variable and braking torque can be modeled as a second-order linear system

MILANS et al.: COMPARING FUZZY AND INTELLIGENT PI CONTROLLERS IN STOP-AND-GO MANOEUVRES

773

Fig. 4. Speed control surface. Fig. 3. Membership functions.

TABLE II RULE BASE

by dening errors of these values with respect to actual values obtained from the vehicles sensors. For tuning, simulation with the car model is used to adjust the membership functions to the actual vehicle that will subsequently be used in the experimental phase. The controller can be easily readjusted to suit any kind of vehicle by slightly modifying the membership functions, in particular, by searching for the optimal tradeoff between good reference tracking and smooth control actions. The output generated is the action on the longitudinal actuatorsi.e., throttle and brake pedals. In particular, the fuzzy output variable Pedal determines which actuator has to be pressed, and the magnitude of the action. The shape of its membership function is dened in terms of Sugeno singletons. The possible output values lie within the range , where indicates the brake pedal is completely depressed and 1 indicates the maximum action is applied to the accelerator pedal . In the present case, these values will be less than unity since the controller will be used at very low speeds (see the bottom plot in Fig. 3). Fig. 4 shows the control surface for the output variable as a function of the input variables of our fuzzy controller. The output values assigned for each rule are listed in Table II.

and are two constant parameters, which where do not necessarily represent physical magnitudes, and whose choice is based on the following guidelines: is usually 1 or 2, and may, although not necessarily, rep resent the system order. should allow and to be of the same order of mag nitude. The term , which is a sort of nonlinear black box identier [17], is computed from the input value at the preceding sample and the th derivative estimate of the output time at the current sample time (8) Using the formalism introduced in (7) and (5) for the two vehicles, the relative velocity dynamics can be expressed as (9)

C. i-PI Controller Intelligent PI controllers are used in this work because they combine the well-known PI structure with an intelligent term that compensates the effects of nonlinear dynamics, disturbances, or uncertainties in the parameters. As a consequence, the nonlinear dynamics of the car at low speeds become controllable with this novel technique. Following [17], [23], a nite-dimensional nonlinear system can be written locally as

where are respectively the engine and brake control variables. If (9) is inverted and merged with a PI controller [20], the resulting i-PI control law is

(10) where are PI gains. Even though this system is nonlinear and has varying parameters, the linearized model of (5) can be useful to tune the PI controller classically (see [20]). and are chosen (see Table III) to Two extra parameters enhance the dynamic behavior and disturbance rejection of the

(7)

774

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 20, NO. 3, MAY 2012

TABLE III CONTROL PARAMETERS

closed-loop system, particularizing (10) in our case to control the throttle

(11)

Fig. 5. Leading vehicle speed and road slope proles.

and to do likewise with the brake

(12)

where is a velocity derivative estimate. Since measurements of the acceleration are available from the CAN bus, this second derivative estimate will not be necessary in this case. Finally, a decision rule will be established to determine whether brake or throttle actions are needed. The control law (12) will be triggered if the reference acceleration is negative and the inter-distance error is less than a xed value . In any other case, the throttle control law (11) will be used. The resulting control law turns the automated vehicle into a nonlinear hybrid system. It is well known how difcult it is to assess the stability of such complex systems. However, a recent work addressing this issue [24] has demonstrated the generalized applicability of this kind of controller to such hybrid systems Eq. (11), Eq. (12), if otherwise if otherwise.

accomplish their control goals while rejecting disturbances induced by the slope of the roadway depicted in Fig. 5. The leading and trailing cars will initially be separated by a distance of 49 m and running at the same speed, 11 ms . The inter-distance dynamic model is parameterized to provide a maximum speed 50 km/h , a maximum 2 ms , a maximum jerk 5 ms , acceleration 6 m. and a minimum inter-distance The controllers will be quantitatively tuned and evaluated with respect to an optimization criterion. Several tness functions are commonly used to tune PI controllers. Among them, the integral absolute error (IAE) is the best adapted to our situation because of its sensitivity. Thus, a cost function including the distance error, the speed error, and the control actions smoothness will be used to design our PID-based controllers

where

1) Tuning Procedure: To evaluate the closed-loop systems behavior with the two controllers, the vehicles dynamics will be simulated, as mentioned above, with a realistic model which takes into account tyres, brakes, and engine dynamics. Measurement noise in the velocity and acceleration CAN-based sensors will also be considered. These corrupting noises will be modeled as additive white Gaussian variables [see the inset in Fig. 6(a)]. The value taken for the transmission rate will be 5 Hz, set by the GPS receiver as detailed in Section IV-A. Fig. 5 shows a velocity prole plot for the leading vehicle. The idea behind this scenario was to evaluate the control algorithms over a wide operating range, trying to emulate coping with the most demanding manoeuvres in urban driving conditions. A robust longitudinal control algorithm [25] will be applied to the rst car to track the setpoints as precisely as possible. Furthermore, both the leading and the trailing cars will have to

is the distance tracking error, is the relative velocity error, and measures the control smoothand ness. The sum of engine and brake control variables is equivalent to the Pedal variable in the fuzzy control implementation. The optimization process gives as a result the PI parameters listed in Table III. Note that i-PI parameters are chosen such that PI gains are the same than for the optimal PI controller and the parameter, for both brake and throttle, was manually tuned thereafter. D. Simulation Results The evolution of the vehicles speeds, the distance between them, the error with respect to the dynamic inter-distance model, acceleration, jerk, and the control action are shown in Fig. 6, which thus provides a synthesis of the most important aspects of the controllers behavior. In this gure and henceforth, the dashdotted red line will be used to correspond to the fuzzy controller, the dashed green line to the i-PI controller, and the solid blue line to the PI controller. One appreciates that the PI controller presents the poorest behavior. With respect to the i-PI and fuzzy controllers, at rst sight their behavior is satisfactory, but there are some important

MILANS et al.: COMPARING FUZZY AND INTELLIGENT PI CONTROLLERS IN STOP-AND-GO MANOEUVRES

775

TABLE IV SIMULATION PERFORMANCE CRITERIA

comfort indicator in Table IVwhich shows that the i-PIs control action is smoother than the fuzzy controllersis coherent with the uctuating behavior of the fuzzy controller observed in the bottom plot of Fig. 6. Both the acceleration and the jerk 2 ms 5 ms ) are satised by constraints ( all three controllers. E. Robustness Analysis Since no parameters appear explicitly in the closed form controller, classical robust control tools cannot here be exploited to analyze the closed loop systems sensitivity to disturbances or to parameter uncertainty. Non-deterministic techniques, in particular Monte Carlo methods, seem to be the most suitable tool to test robustness when model-free control laws are used. To account for parameter uncertainty, we dene the models parameters as distributions of values rather than as single xed values. We then perform a Monte Carlo simulation, running the model repeatedly with 1000 random combinations of parameter values. The parameters in Table I will be considered to follow centred normal distributions of the form . The amplitudes and frequencies of disturbances due to sinusoidal slopes are dened as uniform distributions beto tween their maximum and minimum values ensure that a wide design space is covered. Table V summarizes the Monte Carlo results. One observes that all the implemented controllers are stable within the predened uncertainty domain. However, there are signicant differences between the three approaches. The smoothest action is provided by the PI controller. But with a slightly greater control action activity i-PI tracks distance and relative velocity much better than the optimized PI controller. The fuzzy controller also yields good tracking under the nominal conditions (see Table IV), but is more sensitive to disturbances and parameter uncertainty than the i-PI control law. Fig. 7 shows the inter-distance and speed errors in the worst case for all three control strategies. One observes that both the fuzzy and the i-PI controllers improve the inter-distance and relative velocity tracking results obtained with the optimized PI. It also conrms that i-PIs behavior is remarkably good for both distance and relative velocity. IV. EVALUATION OF THE CONTROLLERS To validate the proposed control algorithms, the two controllers were implemented in the AUTOPIA program control architecture for autonomous vehicles [26]. This section presents a brief description of the real cars used for the experimental phase and their automation process. Then the real results at the CARs facilities will be described.

Fig. 6. (a) Leading and trailing vehicle speeds. (b) Inter-vehicular distance. (c) Inter-distance error. (d) Acceleration. (e) Jerk. (f) Normalized control action.

differences in terms of safety, comfort, and rejection of disturbances. Concerning the inter-distance error relative to the dynamic reference, all the controllers present an initial transient that is corrected after approximately 5 seconds. From then on, i-PI provides more precise tracking than the fuzzy controller (see Table IV). Even though the fuzzy controllers tracking quality is not as good as i-PIs, the vehicle rarely exceeds the safe reference distance. The PI controller presents the worst behaviour, with inter-distance errors greater than 3 m. Also the values of the

776

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 20, NO. 3, MAY 2012

TABLE V MONTE CARLO ROBUSTNESS ANALYSIS

Fig. 7. Worst case in Monte Carlo simulations for all three controllers. (a) Position error. (b) Speed error.

mation, providing 1-cm precision. An inertial measurement unit (IMU) is installed in the convertible car to provide positioning in case of GPS receiver failure [29]. A Personal Computer Memory Card International Association (PCMCIA) Proxim Wireless ComboCard is installed in the PC of each car, and a central station is used to send the relevant information from the leading to the trailing car [30]. The trailing vehicle is equipped with an industrial on-board PC that is in charge of receiving the information coming from the wireless communication system and the sensorial inputs, and of sending the output generated to the actuators in each control cycle (200 ms). Remark 1: The system has been tested using DGPS in place of RTK-DGPS receivers, without degradation of performance. This kind of receiver removes the dependence on a local station to transmit the differential corrections, so that the system can work over hundreds of kilometers. Trials with low-cost commercial GPS receivers showed that they are as yet inappropriate for this kind of application. B. Real Results Several trials were conducted at the CARs private driving circuit using the experimental vehicles. This circuit represents an inner-city area, with a combination of straight-road segments, bends, and different road slopes. During these trials, a tuning renement was applied to the controllers because of the complexity of the translation from simulation to the real world. To compare the two controllers in conditions as equal as possible, a predened route was recorded. This route was rst traveled over with the manually driven vehicle, and all the relevant variables to perform the controlposition, speed and accelerationwere stored. In this way, the human inuence in two consecutive trials was removed. In a parallel line, a PI controller previously developed to perform this application [5] was used to compare the novel controllers with previous results. Fig. 9. shows the results for each of the controllersPI, i-PI, and fuzzyduring this experiment. The distance between vehicles at the beginning of the test was set at 6 metres. Once this distance was achieved with 1-cm accuracy using the RTK-DGPS positioning system, the test was initiated. The top plot depicts the trailing vehicles speed with respect to the leading one. The second from the top shows the desired inter-distance and the values obtained using the designed controllers. The third shows the desired relative velocity and the values obtained by the controllers. The fourth and fth show the acceleration and the jerk, respectively, for the three controllers. The bottom plot shows the action on the accelerator and brake pedals, with the values being , where 1 indicates an action on normalized to the range on the brake. the accelerator pedal and At the beginning, the leading car is stopped at the predened safety distance6 m. During the rst minute, the fuzzy

Fig. 8. Commercial prototype vehicles used for the experimental phase.

A. Experimental Vehicles Two vehicles were used for the experimental phase: a fully-automated vehicle and a manually driven one. The former is a convertible Citron C3 Pluriel (see Fig. 8). The car is equipped with automatic driving capabilities with hardware modications made to the throttle and the brake pedal actions. The latter vehicle is an electric Citron Berlingo van (see Fig. 8) also equipped with automatic driving capabilities. For the purpose of this work, it was driven by a human driver making the leading cars behavior as close to a real trafc situation as possible. With respect to the automation process, the Pluriels throttle is controlled by an analogue signal that represents the pressure on the pedal, generated with an analogue card [27]. For the brake, an electro-hydraulic braking system is mounted in parallel with the original one [28], and is controlled via an I/O digital-analogue CAN card. Both vehicles are equipped with real time kinematicdifferential global positioning systems (RTK-DGPS) working at 5 Hz as the main sensor. This sensor is used to acquire driving infor-

MILANS et al.: COMPARING FUZZY AND INTELLIGENT PI CONTROLLERS IN STOP-AND-GO MANOEUVRES

777

TABLE VI EXPERIMENTATION PERFORMANCE CRITERIA

the desired inter-distance. Later, the fuzzy controller follows the reference with better precision than the i-PI up to second 130 when the leading car is stopped. Then, the i-PI controller reaches the predened safety distance faster than the fuzzy controller. Variations in the PI controller are greater than the other two controllers throughout the test. Concerning the acceleration and jerk, all three controllers sat2 ms 5 ised the initial prerequisites ms . With respect to the control action, the smoothest is attained by the i-PI controller. The fuzzy controller presents the greatest oscillations in control actions, but they can be neglected from a perspective of the comfort of the cars occupants. To quantify these results, the control quality indicators introduced in Section III-C1 are compared in Table VI. One observes that the tracking error is better in the fuzzy than in the i-PI controller, but that the control action is harsher in the fuzzy controller. In brief therefore, the fuzzy controller presents as its main advantage that it can be intuitively retuned and behaves slightly better in tracking, the i-PI controller has smoother control action, and the PI controller presents the poorest behavior of the three. Remark 2: There are numerous effects that appear in real experiments that are extremely difcult to quantify (weather conditions, the pavement of the road, pressure on the wheel, etc.). Thus, although the i-PI controller worked better in our simulations, the fuzzy controller parameters were less sensitive than those of the i-PI when translated to the real world. V. CONCLUSION Stop-and-go manuvres constitute one of the most important and as yet unsolved topics in the automotive sector. This paper has presented two control techniquesfuzzy logic and i-PI controllerswith which to solve this problem. First, a car model was used to validate the proposed controllers. Then, two real cars were used to check the behaviour of the control algorithm in real circumstances. The following lessons can be learned from this work. The fuzzy controller developed can be easily adapted from simulations to the real word since it is based on human behavior and is a model-free control technique. The i-PI controller, that it is only based on a rough model of the system, provides smoother action on the throttle and brake pedals, thus increasing the vehicle occupants comfort. The behavior of both these new controllers is signicantly better than a controller developed previously [5] to perform this manoeuvre. The proposed control techniques can signicantly reduce trafc jams, making driving easier by relieving the human driver of some tedious tasks.

Fig. 9. (a) Leading and trailing vehicle speeds. (b) Distance between vehicles. (c) Relative velocity. (d) Acceleration. (e) Jerk. (f) Normalized control action.

controller is very close to the reference inter-distance and the i-PI controller is slightly under the desired value. One can appreciate how the PI controller presents the poorest behavior with greater values than the reference inter-distance. This behavior is reected in the desired relative velocity. The greatest inter-distance error occurs at around second 64, because the autonomous vehicle is driving around a curved stretch. Then, at around second 85, the leading cars speed is signicantly slowed to close to zero at the end of a straight stretch. When the leading vehicle starts again, a minimal delay appears in the fuzzy controller that is resolved with hard action on the throttle to recover

778

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 20, NO. 3, MAY 2012

The promising results obtained with the two controllers in the work described in this paper will be pursued analytically and experimentally with more vehicles and in other advanced driver assistance systems. REFERENCES
[1] A. Kestinga, M. Treibera, M. Schnhofa, and D. Helbing, Adaptive cruise control design for active congestion avoidance, Transportation Res. Pt. C, Emerging Technol., vol. 16, no. 6, pp. 668683, Dec. 2008. [2] B. van Arem, C. J. G. van Driel, and R. Visser, The impact of cooperative adaptive cruise control on trafc-ow characteristics, IEEE Trans. Intell. Transportation Syst., vol. 7, no. 4, pp. 429436, 2006. [3] N. B. Hounsell, B. P. Shrestha, J. Piao, and M. McDonald, Review of urban trafc management and the impacts of new vehicle technologies, IET Intell. Transport Syst., vol. 3, no. 4, pp. 419428, 2009. [4] S. Moon, I. Moon, and K. Yi, Design, tuning and evaluation of a fullrange adaptive cruise control system with collision avoidance, Control Eng. Pract., vol. 17, pp. 442455, 2009. [5] J. Martinez and C. C. de Wit, A safe longitudinal control for adaptive cruise control and stop-and-go scenarios, IEEE Trans. Control Syst. Technol., vol. 15, no. 1, pp. 246258, Jan. 2007. [6] D. Swaroop, J. Hedrick, C. Chien, and P. Ioannou, A comparision of spacing and headway control laws for automatically controlled vehicles 1, Veh. Syst. Dyn., vol. 23, no. 1, pp. 597625, 1994. [7] J. Naranjo, C. Gonzlez, R. Garcia, and T. de Pedro, ACC Stop&go maneuvers with throttle and brake fuzzy control, IEEE Trans. Intell. Transportation Syst., vol. 7, no. 2, pp. 213225, Jun. 2006. [8] S. Dermann and R. Isermann, Nonlinear distance and cruise control for passenger cars, in Proc. Amer. Control Conf., 1995, pp. 30813085. [9] J. Gerdes and J. Hedrick, Vehicle speed and spacing control via coordinated throttle and brake actuation, Control Eng. Pract., vol. 5, no. 11, pp. 16071614, 1997. [10] L. Nouveliere and S. Mammar, Experimental vehicle longitudinal control using a second order sliding mode technique, Control Eng. Pract., vol. 15, no. 8, pp. 943954, 2007. [11] H. Raza, Z. Xu, B. Yang, and P. Ioannou, Modeling and control design for a computer-controlled brake system, IEEE Trans. Control Syst. Technol., vol. 5, no. 3, pp. 279296, Sep. 1997. [12] K. Yi and J. Chung, Nonlinear brake control for vehicle CW/CA systems, IEEE/ASME Trans. Mechatron., vol. 6, no. 1, pp. 1725, Jan. 2001. [13] Y. Dote and S. J. Ovaska, Industrial applications of soft computing: A review, Proc. IEEE, vol. 89, no. 9, pp. 12431265, Sep. 2001. [14] L. Zadeh, Fuzzy sets, Inf. Control, vol. 8, pp. 338353, 1965.

[15] M. Sugeno and M. Nishida, Fuzzy control of a model car, Fuzzy Sets Syst., vol. 16, no. 2, pp. 103113, 1985. [16] M. Fliess and J. C. , Intelligent PID controllers, in Proc. 16th Mediterranean Conf. Control Autom., 2008, pp. 326331. [17] M. Fliess and C. Join, Model-free control and intelligent PID controllers: Towards a possible trivialization of nonlinear control?, presented at the 15th IFAC Symp. Syst. Identication (SYSID), SaintMalo, France, 2009. [18] R. Rajamani, Vehicle Dynamics and Control.. New York: Springer, 2005. [19] H. Pacejka and E. Bakker, The magic formula tyre model, Veh. Syst. Dyn., vol. 21, pp. 118, 2004. [20] K. Astrm and T. Hgglund, Advanced PID Controllers, Instrument Soc. Amer., 2006, TJ 223.P55A85. [21] V. Milans, E. Onieva, J. Prez, T. de Pedro, and C. Gonzlez, Control de velocidad adaptativo para entornos urbanos congestionados, Revista Iberoamericana Autom. Inf. Ind., vol. 6, no. 4, pp. 6673, Oct. 2009. [22] M. Sugeno and M. Nishida, Fuzzy control of model car, Fuzzy Sets Syst., vol. 16, pp. 103113, 1985. [23] J. Villagra, V. Milans, J. Prez, and T. de Pedro, Control basado en PID inteligentes: Aplicacin al control robusto de velocidad en entornos urbanos, Revista Iberoamericana Autom. Inf. Ind., vol. 7, no. 4, pp. 4452, Oct. 2010. [24] R. Bourdais, M. Fliess, C. Join, and W. Perruquetti, Towards a modelfree output tracking of switched nonlinear systems, presented at the IFAC Symp. Nonlinear Control Syst., Pretoria, South Africa, 2007. [25] V. Milans, J. Villagr, J. Prez, and C. Gonzlez, Low-speed longitudinal controllers for mass-produced cars: A comparative study, IEEE Trans. Ind. Electron., to be published. [26] J. Prez, C. Gonzlez, V. Milans, E. Onieva, J. Godoy, and T. de Pedro, Modularity, adaptability and evolution in the AUTOPIA architecture for control of autonomous vehicles, in Proc. IEEE Int. Conf. Mechatron. (ICM), 2009, pp. 15. [27] V. Milans, D. Llorca, B. Vinagre, C. Gonzlez, and M. Sotelo, Clavileo: Evolution of an autonomous car, in Proc. 13th Int. IEEE Conf. Intell. Transportation Syst., 2010, pp. 11291134. [28] V. Milans, C. Gonzlez, J. Naranjo, E. Onieva, and T. De Pedro, Electro-hydraulic braking system for autonomous vehicles, Int. J. Autom. Technol., vol. 1, no. 11, pp. 8995, Feb. 2010. [29] V. Milans, J. E. Naranjo, C. Gonzalez, J. Alonso, and T. de Pedro, Autonomous vehicle based in cooperative gps and inertial systems, Robotica, vol. 26, pp. 627633, 2008. [30] V. Milans, J. Godoy, J. Prez, B. Vinagre, C. Gonzlez, E. Onieva, and A. J. , V2I-based architecture for information exchange among vehicles, presented at the 7th Symp. Intell. Autonomous Veh., Lecce, Italy, 2010.